Multimodal hybrid linear auto-weighting models: Application of ultraviolet spectroscopy for growth prediction of marine pathogenic bacteria

Abstract

Spectral contactless detection techniques are real-time and extremely sensitive. It shows great potential for rapid prediction of growth trends of marine microorganisms. This paper proposes a real-time detection system that integrates ultraviolet spectroscopy (UV) technology with the PPSA model (1DCNN-PLSR Parallel 1DCNN-SVR Adaptation). The system realized the detection of microbial concentration and the prediction of growth trend. A UV absorption spectral dataset was constructed by collecting ultraviolet (UV) spectral data of actinomycetes during growth and combining it with the microscopic hemocyte plate counting method. For multiple feature intervals characterizing actinomycetes in UV absorption spectral curves, spectral data feature extraction was performed by a one-dimensional convolutional neural network (1DCNN), and prediction experiments were conducted using multiple parallel network models. Further exploration of weight adaptive fine-tuning of model parameter shares to optimize overall prediction accuracy. The experimental evaluation showed that the model achieved an R² of 0.9994, an MAE of 0.4181, and an average error of 2.0 * 10⁶cells/mL in the assay results. Rapid characterization and detection of the growth of known and unknown pathogenic bacteria using ultraviolet absorption spectroscopy is valuable for the study of marine and human ecosystems.